This code develops a multi-task learning neural combinatorial optimization (NCO) model for cross-problem generalization for routing problems.

A comparison of gaps on eleven VRPs (Left: box plot, Right: radar plot). ST represents the unified model trained with single-task learning on CVRP, ST_all represents the unified model with single-task learning on OVRPBLTW, and MT represents our approach, i.e., the unified model with multi-task learning on five VRPs. ST_FT and MT_FT are the fine-tuning models

Please cite the paper https://arxiv.org/abs/2402.16891, if you find the code helpful,

@article{liu2024multi, \ title={Multi-Task Learning for Routing Problem with Cross-Problem Zero-Shot Generalization}, \ author={Liu, Fei and Lin, Xi and Zhang, Qingfu and Tong, Xialiang and Yuan, Mingxuan}, \ journal={International Conference on Knowledge Discovery and Data Mining (KDD)}, \ year={2024} \ }

Files in MTNCO

• MTPOMO: the implementation of multi-task learning with attribute composition based on POMO.
• Trained_models: the pre-trained unified models with problem sizes 50 and 100
• Test_instances: test instances of 11 VRPs
• utils: utils

Train & Test

cd MTPOMO/POMO/

Train: python train_n50.py

Test: python test_n50.py

Acknowledgements

Our implementation is based on the code of POMO. Thanks to them.